In 1974, Peter Szolovits predicted: By the 1980s, most enormous hospitals would adopt electronic health records. Although the technology has not developed as quickly as expected, the U.S. government is now making a major effort to move hospitals from paper files filled with notes to a secure and capable electronic system for collecting, storing, and retrieving medical records. Today, Szolovits—a professor of computer science and engineering, health sciences, and technology and leader of the Clinical Decision Making Group at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL)—is at the forefront of a movement to make health information technology more effective and useful to both providers and patients.
Szolovits’ current focus is on using natural language processing to better understand medical records, which often contain unstructured narrative notes and abbreviations. As part of the government’s push for electronic health records, Szolovits is a participant in one of four Strategic Health IT Advanced Research Projects (SHARP) from the Office of the National Coordinator for Health IT. The project, led by colleagues at the Mayo Clinic, focuses on the secondary operate of electronic health records, providing a way to analyze clinical data sets for research, quality assessment, and public health support. The main technical challenges include developing natural language processing methods that enable computer programs to automatically extract clinical files, events, and relationships from narrative text in the records, and to combine the extracted data with existing information from laboratory tests and physician orders to identify patient conditions and current treatments.
“My interest in natural language processing was rekindled about 12 years ago when I noticed that a lot of important medical data was actually locked up in these narrative notes, and we needed to have some way to dig them out so we could use them,” Szolovits says.
William Long, a principal research scientist at Clinical Decision Making Group who has worked with Szolovits for more than 30 years, has developed several programs that scan electronic nursing reports and intensive care unit (ICU) discharge papers, looking for key words and phrases that provide clues about a patient’s condition. Drawing on information gleaned from the Uniform Medical Language System, a compilation of more than 150 medical dictionaries, Clinical Decision Making Group has programmed the system to identify a comprehensive list of terms and key concepts. The program can then provide physicians with a brief summary of the compiled information.
Until now, the technology has been used more for clinical trials than for diagnostic purposes. For example, it was used to gather a group of patients who had the same disease but responded differently to identical treatments. With systems that can organize and analyze medical records, doctors can discover whether genetics, external medications, or personal habits have affected a patient, and learn more about which treatments are most effective.
Szolovits still believes in using AI in the diagnostic process, but in a different way than he originally imagined. He and his colleagues—Roger Mark; George Verghese; and colleagues from CSAIL, the Laboratory for Information and Decision Systems, Harvard-MIT Health Sciences and Technology, and Beth Israel Deaconess Medical Center—collected data on about 35,000 ICU admissions at a enormous Boston hospital. CSAIL graduate student Caleb Hug used the data to create predictive models that, for each significant change in a patient’s condition, estimate how well they’re likely to fare in the future. Such acuity models can warn doctors of danger and are also useful in determining the resources needed to aid a particular patient.
The same methodology can also be used for more detailed predictions. Hug’s research used the technique to predict when it would be secure to wean patients off life-support devices such as ventilators and intra-aortic pumps, as well as whether a patient was likely to develop septic shock, hypotension, or kidney failure.
Outside the ICU, the group is tackling the challenge of recording doctor-patient interactions and translating those conversations into actionable information. For a project at Children’s Hospital Boston, group members recorded about 100 encounters at the Pediatric Environmental Health Center, where doctors often see cases of lead poisoning in children. Each doctor-patient interaction is recorded, translated into English text using a speech-understanding program, and then analyzed for key terms and phrases using a natural-language processing program developed by assistant professor of computer science and engineering Regina Barzilay and her group.
Although the project has proven challenging due to the difficulty of building a high-fidelity speech-understanding component, Szolovits believes it could improve medical visits for doctors. In the future, Szolovits says, the technology could also be used by hospital nurses so they can focus more on patient care than taking notes.
Despite the challenges of moving to more technologically advanced systems, Szolovits and Long believe that natural language processing programs like those developed by the Clinical Decision Making Group could significantly improve the quality of medical care.
“I think it’s going to revolutionize healthcare. It’s going to be much easier to get all of this information in a form that we can actually do something with and process it in a way that benefits everyone,” Long says. “Doctors will benefit from being able to look at years of patient care and turn it into research about what’s working, what’s not working, how to improve care, how to treat patients better.”
For more information about the Clinical Decision Making Group, please visit: http://groups.csail.mit.edu/medg/.